The leak is considered as purely negative and damaging in this competition as I can see. I would like to suggest alternative view to the leak that enables learning something new from it, so it is not waste of time anymore.
\n\nWhen a host and Kaggle prepare competition the real data are transformed, for example anonymized, obfuscated and filtered. Many questions about data and how it is collected and prepared usually remain unanswered. The leak in this competition gives us a unique chance to see what is hidden. We can compare real data and competition data.\nWe can try find out why city and country sometimes mismatch. Is it data property or some kind of obfuscation?
\n\nDoes demo data have the same number of \"(not set)\" and \"not available in demo dataset\" values or they were introduced artificially?
\n\nOne thing I would like share shows how far real data can be from one we use in the competition. Comparing demo account and train data I found the following situation. A user spent almost 3 hours buying stuff for about $9000, he/she made several purchases during this single session with total 400 page views. Corresponding record from competition train has totals.transactionRevenue equals to $1700, but the same number of pageviews/hits. It turned out that only last purchase of the user within single session is in train data. I saw this for few users and cannot state it affects all multi purchase sessions. I think this issue has nothing to do with Kaggle, as it is based on Google Analytics Sample Dataset for BigQuery, but we do not know if this issue exists for test as this is not part of Sample Dataset.
\n\nI believe kagglers can discover much more interesting if dig deeper.
", "messages": [ { "id": "403495", "postDate": "10/13/2018 18:45:57", "content": "The leak is considered as purely negative and damaging in this competition as I can see. I would like to suggest alternative view to the leak that enables learning something new from it, so it is not waste of time anymore.
\n\nWhen a host and Kaggle prepare competition the real data are transformed, for example anonymized, obfuscated and filtered. Many questions about data and how it is collected and prepared usually remain unanswered. The leak in this competition gives us a unique chance to see what is hidden. We can compare real data and competition data.\nWe can try find out why city and country sometimes mismatch. Is it data property or some kind of obfuscation?
\n\nDoes demo data have the same number of \"(not set)\" and \"not available in demo dataset\" values or they were introduced artificially?
\n\nOne thing I would like share shows how far real data can be from one we use in the competition. Comparing demo account and train data I found the following situation. A user spent almost 3 hours buying stuff for about $9000, he/she made several purchases during this single session with total 400 page views. Corresponding record from competition train has totals.transactionRevenue equals to $1700, but the same number of pageviews/hits. It turned out that only last purchase of the user within single session is in train data. I saw this for few users and cannot state it affects all multi purchase sessions. I think this issue has nothing to do with Kaggle, as it is based on Google Analytics Sample Dataset for BigQuery, but we do not know if this issue exists for test as this is not part of Sample Dataset.
\n\nI believe kagglers can discover much more interesting if dig deeper.
", "rawMarkdown": "The leak is considered as purely negative and damaging in this competition as I can see. I would like to suggest alternative view to the leak that enables learning something new from it, so it is not waste of time anymore.\n\nWhen a host and Kaggle prepare competition the real data are transformed, for example anonymized, obfuscated and filtered. Many questions about data and how it is collected and prepared usually remain unanswered. The leak in this competition gives us a unique chance to see what is hidden. We can compare real data and competition data.\nWe can try find out why city and country sometimes mismatch. Is it data property or some kind of obfuscation? \n\nDoes demo data have the same number of \"(not set)\" and \"not available in demo dataset\" values or they were introduced artificially?\n\nOne thing I would like share shows how far real data can be from one we use in the competition. Comparing demo account and train data I found the following situation. A user spent almost 3 hours buying stuff for about $9000, he/she made several purchases during this single session with total 400 page views. Corresponding record from competition train has totals.transactionRevenue equals to $1700, but the same number of pageviews/hits. It turned out that only last purchase of the user within single session is in train data. I saw this for few users and cannot state it affects all multi purchase sessions. I think this issue has nothing to do with Kaggle, as it is based on Google Analytics Sample Dataset for BigQuery, but we do not know if this issue exists for test as this is not part of Sample Dataset.\n\nI believe kagglers can discover much more interesting if dig deeper.", "votes": null }, { "id": "405282", "postDate": "10/17/2018 07:13:47", "content": "Kaggle found a good way to avoid leak usage and save this competition. One interesting result of Kaggle decision is that new train and test data can be validated using demo account, so in case of any error it would become evident. \nThus this competition is going to be most transparent one.
", "rawMarkdown": "Kaggle found a good way to avoid leak usage and save this competition. One interesting result of Kaggle decision is that new train and test data can be validated using demo account, so in case of any error it would become evident. \nThus this competition is going to be most transparent one.", "votes": null }, { "id": "405348", "postDate": "10/17/2018 09:42:47", "content": "That's really interesting idea.\nThank you @Alexander Firsov
", "rawMarkdown": "That's really interesting idea.\nThank you @Alexander Firsov", "votes": null }, { "id": "405419", "postDate": "10/17/2018 13:01:33", "content": "No duplicate sessionIDs? That was also something that I was wondering about.
", "rawMarkdown": "No duplicate sessionIDs? That was also something that I was wondering about.", "votes": null }, { "id": "405546", "postDate": "10/17/2018 17:20:51", "content": "Erik, duplicate session id I believe is how GA works.\nFor example session from train 2571951630476198714_1472105745 is split between two days. Each part has purchases. In GA interface this is Client Id: 1949416477.1472105721 / BigQuery Client ID: \n8372680016470641913.\nThe user has the same two sessions in GA interface.
\n\nP.S. what is interesting is that second part of the session has different geoNetwork.country and geoNetwork.networkDomain, user jumped from Venezuela to US.
", "rawMarkdown": "Erik, duplicate session id I believe is how GA works.\nFor example session from train 2571951630476198714_1472105745 is split between two days. Each part has purchases. In GA interface this is Client Id: 1949416477.1472105721 / BigQuery Client ID: \n8372680016470641913.\nThe user has the same two sessions in GA interface.\n\nP.S. what is interesting is that second part of the session has different geoNetwork.country and geoNetwork.networkDomain, user jumped from Venezuela to US.", "votes": null } ], "comments": [ { "id": 405282, "author_name": "alexfir", "author_url": "", "post_date": "10/17/2018 07:13:47", "content": "Kaggle found a good way to avoid leak usage and save this competition. One interesting result of Kaggle decision is that new train and test data can be validated using demo account, so in case of any error it would become evident. \nThus this competition is going to be most transparent one.
", "votes": null, "replies": [] }, { "id": 405348, "author_name": "kaggle2echo", "author_url": "", "post_date": "10/17/2018 09:42:47", "content": "That's really interesting idea.\nThank you @Alexander Firsov
", "votes": null, "replies": [] }, { "id": 405419, "author_name": "erikbruin", "author_url": "", "post_date": "10/17/2018 13:01:33", "content": "No duplicate sessionIDs? That was also something that I was wondering about.
", "votes": null, "replies": [ { "id": 405546, "author_name": "alexfir", "author_url": "", "post_date": "10/17/2018 17:20:51", "content": "Erik, duplicate session id I believe is how GA works.\nFor example session from train 2571951630476198714_1472105745 is split between two days. Each part has purchases. In GA interface this is Client Id: 1949416477.1472105721 / BigQuery Client ID: \n8372680016470641913.\nThe user has the same two sessions in GA interface.
\n\nP.S. what is interesting is that second part of the session has different geoNetwork.country and geoNetwork.networkDomain, user jumped from Venezuela to US.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "403495": "The leak is considered as purely negative and damaging in this competition as I can see. I would like to suggest alternative view to the leak that enables learning something new from it, so it is not waste of time anymore.\n\nWhen a host and Kaggle prepare competition the real data are transformed, for example anonymized, obfuscated and filtered. Many questions about data and how it is collected and prepared usually remain unanswered. The leak in this competition gives us a unique chance to see what is hidden. We can compare real data and competition data.\nWe can try find out why city and country sometimes mismatch. Is it data property or some kind of obfuscation? \n\nDoes demo data have the same number of \"(not set)\" and \"not available in demo dataset\" values or they were introduced artificially?\n\nOne thing I would like share shows how far real data can be from one we use in the competition. Comparing demo account and train data I found the following situation. A user spent almost 3 hours buying stuff for about $9000, he/she made several purchases during this single session with total 400 page views. Corresponding record from competition train has totals.transactionRevenue equals to $1700, but the same number of pageviews/hits. It turned out that only last purchase of the user within single session is in train data. I saw this for few users and cannot state it affects all multi purchase sessions. I think this issue has nothing to do with Kaggle, as it is based on Google Analytics Sample Dataset for BigQuery, but we do not know if this issue exists for test as this is not part of Sample Dataset.\n\nI believe kagglers can discover much more interesting if dig deeper.", "405282": "Kaggle found a good way to avoid leak usage and save this competition. One interesting result of Kaggle decision is that new train and test data can be validated using demo account, so in case of any error it would become evident. \nThus this competition is going to be most transparent one.", "405348": "That's really interesting idea.\nThank you @Alexander Firsov", "405419": "No duplicate sessionIDs? That was also something that I was wondering about.", "405546": "Erik, duplicate session id I believe is how GA works.\nFor example session from train 2571951630476198714_1472105745 is split between two days. Each part has purchases. In GA interface this is Client Id: 1949416477.1472105721 / BigQuery Client ID: \n8372680016470641913.\nThe user has the same two sessions in GA interface.\n\nP.S. what is interesting is that second part of the session has different geoNetwork.country and geoNetwork.networkDomain, user jumped from Venezuela to US." }, "source": "meta" }